prompt-atlas-ecl

Chapter 5 · Quantum Bridges and Cosmic Noise — Expansion

Canonical: PROMPT_ATLAS.md#ch5-quantum-bridges-and-cosmic-noise · Prompts: prompts/ch05.yaml · Part: III

Worked Example — Cosmic Pattern Finder, Made Honest

Original: “Train AI to search cosmic background radiation for patterns dismissed as noise — what anomalies emerge?”

1. Pre-register the search — Specify (a) the noise channel, (b) the anomaly metric, (c) the false-positive budget, before running.
2. Use geometric distances — Compare predicted vs. observed distributions with Sinkhorn-Wasserstein or MMD; raw mean-squared error hides multimodal anomalies.
3. Hold out a sky patch — Train on the rest, evaluate on the held-out patch. Anomalies that don’t generalize are noise rediscovered.
4. Force interpretability — Each anomaly must have a story (mechanism, equation, or testable prediction) before being elevated.
5. Publish nulls. A million null searches matter as much as one positive.

Prompt Templates

# Pre-registered anomaly search
"Define an anomaly search over .
 Required fields: noise model, anomaly metric, false-positive budget,
 hold-out strategy, interpretability requirement, null publication plan."

# Bridge proposal
"Propose three families of partial bridges between  and .
 For each, state the regime where it works, the regime where it breaks,
 and one experiment that could distinguish them within 50 years."

# Ethics-of-discovery
"If a model produces a prediction that humans cannot mechanistically explain
 but that consistently outperforms theory across  domains,
 what minimum interpretability bar must be met before action is taken?"

Anti-patterns

• p-hacking with bigger search spaces. AI multiplies the number of comparisons; without pre-registration this guarantees fake discoveries.
• Single-metric loss. A model trained only to minimize MSE will smooth away the very anomalies you want.
• Oracle worship. If a model is right and you don’t know why, that is information about your understanding, not a license to act.
• Hidden hold-outs. Reusing the held-out set silently destroys it.

Try This

1. Pre-Register — Take a hypothesis you care about. Write the pre-registration before running anything.
2. Wasserstein vs. MSE — Compare the same model under sinkhorn_wasserstein and MSE on a multimodal target. Note the anomalies one surfaces and the other hides.
3. Story-Behind-the-Number — For one model output you trust, write the mechanistic story in two sentences. If you can’t, that’s the work.
4. Publish a Null — Write up one negative result you ran this year.
5. Interpretability Budget — For your next experiment, allocate ≥20% of compute to explanation.

Repo Cross-Links

• losses_geom.py — sinkhorn_wasserstein, mmd2, gaussian_kl_sym. These are the mathematical embodiment of “lattice of partial bridges” — distances rather than identities.
• testers/z3_tester.py — formal verifier. Even if Z3 is not installed, the Python fallback runs. Use it to encode invariants the model must not violate.
• tracking/__init__.py — PAE_TRACKING=1 enables MLflow-or-fallback experiment tracking; nulls deserve to be tracked too.

Guide for AI & Humanity

• Interpretability is consent. A society has not agreed to a discovery it cannot understand.
• Publish nulls. Otherwise the literature is folklore weighted toward novelty.
• Honor the noise. Cosmic noise was once “static” until it became the CMB. Listen before erasing.
• Reproducibility before celebration. Track the run before announcing the result.

Citations & Further Reading

• Lee Smolin, The Trouble with Physics (2006).
• Sabine Hossenfelder, Lost in Math (2018).
• Carlo Rovelli, Helgoland (2020).
• Ioannidis, J. P. A. (2005), “Why Most Published Research Findings Are False” — for the discipline of pre-registration.

This site is open source. Improve this page.